pranavrajkumar11-collab/Settlement-Undercount-Project

Testing whether global population datasets undercount dispersed rural settlements, using a fine-tuned geospatial foundation model across five East African countries. Independent undergraduate research. # Settlement Morphology and Rural Population Undercount in East Africa Independent research project quantifying whether global gridded population datasets (WorldPop, GHS-POP) systematically undercount people living in dispersed rural settlements, using a fine-tuned geospatial foundation model across five East African countries: **Kenya, Tanzania, Uganda, Rwanda, and Burundi**. Motivated by Láng-Ritter et al. (2025, *Nature Communications*), which showed gridded population products undercount rural populations globally but did not identify *which kinds* of rural settlements drive the error. This project tests the hypothesis that the undercount concentrates in low building-density ("isolated") settlement morphologies — scattered homesteads that satellite-derived building products struggle to see — relative to clustered villages. ## Key Findings - **Direction-consistent morphology gap in 5/5 countries.** Stratifying rural DHS survey clusters by Google Open Buildings density terciles, gridded population products underrepresent low-density (isolated) settlement areas relative to clustered ones, with country-level bootstrap confidence intervals excluding zero against all three independent reference products. - **Scope of claims.** Absolute bias magnitudes are sensitive to dwelling-area calibration assumptions, so claims are restricted to the direction and consistency of the gap, not effect sizes. ## Pipeline | Notebook | Purpose | |---|---| | `00_setup_auth.ipynb` | Google Earth Engine + HuggingFace authentication, Drive setup | | `01_sentinel_preprocessing.ipynb` | Continental-scale GEE preprocessing: 8-channel Sentinel-1 SAR + Sentinel-2 optical composites (dry + wet season), patch export | | `02_model_finetuning.ipynb` | Prithvi-EO-1.0-100M (NASA/IBM) semantic segmentation fine-tuning: frozen encoder, ~31.5M trainable decoder params, weighted cross-entropy for extreme class imbalance (settlements <1% of pixels); crash-resumable tiled inference | | `03_crossvalida …